Self-Driving Car Engineer Nanodegree¶
Deep Learning¶
Project: Build a Traffic Sign Recognition Classifier¶
In this notebook, a template is provided for you to implement your functionality in stages which is required to successfully complete this project. If additional code is required that cannot be included in the notebook, be sure that the Python code is successfully imported and included in your submission, if necessary. Sections that begin with 'Implementation' in the header indicate where you should begin your implementation for your project. Note that some sections of implementation are optional, and will be marked with 'Optional' in the header.
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. In addition, Markdown cells can be edited by typically double-clicking the cell to enter edit mode.
Step 0: Load The Data¶
# Load pickled data
import pickle
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
import pandas as pd
import random
from PIL import Image
import pickle
import numpy as np
import cv2
import tensorflow as tf
import matplotlib.pyplot as plt
# Visualizations will be shown in the notebook.
%matplotlib inline
# TODO: Fill this in based on where you saved the training and testing data
training_file = "train.p"
testing_file = "test.p"
with open(training_file, mode='rb') as f:
train = pickle.load(f)
with open(testing_file, mode='rb') as f:
test = pickle.load(f)
X_train_orig, y_train_orig = train['features'], train['labels']
X_test, y_test = test['features'], test['labels']
# Visualizations will be shown in the notebook.
%matplotlib inline
sns.set_style("white")
sns.distplot(y_train_orig,kde=False)
data = pd.read_csv("./signnames.csv")
data
#print(X_train_new[0][0][0])
#print(type(X_train_new))
#print(X_train[0].astype(np.float32))
Step 1: Dataset Summary & Exploration¶
The pickled data is a dictionary with 4 key/value pairs:
'features'is a 4D array containing raw pixel data of the traffic sign images, (num examples, width, height, channels).'labels'is a 2D array containing the label/class id of the traffic sign. The filesignnames.csvcontains id -> name mappings for each id.'sizes'is a list containing tuples, (width, height) representing the the original width and height the image.'coords'is a list containing tuples, (x1, y1, x2, y2) representing coordinates of a bounding box around the sign in the image. THESE COORDINATES ASSUME THE ORIGINAL IMAGE. THE PICKLED DATA CONTAINS RESIZED VERSIONS (32 by 32) OF THESE IMAGES
Complete the basic data summary below.
### Replace each question mark with the appropriate value.
X_train = np.copy(X_train_orig)
y_train = np.copy(y_train_orig)
# TODO: Number of training examples
n_train = len(X_train)
# TODO: Number of testing examples.
n_test = len(X_test)
# TODO: What's the shape of an traffic sign image?
image_shape = np.shape(X_train[0])
# TODO: How many unique classes/labels there are in the dataset.
n_classes = len(set(y_train))
print("Number of training examples =", n_train)
print("Number of testing examples =", n_test)
print("Image data shape =", image_shape)
print("Number of classes =", n_classes)
Visualize the German Traffic Signs Dataset using the pickled file(s). This is open ended, suggestions include: plotting traffic sign images, plotting the count of each sign, etc.
The Matplotlib examples and gallery pages are a great resource for doing visualizations in Python.
NOTE: It's recommended you start with something simple first. If you wish to do more, come back to it after you've completed the rest of the sections.
"""
Plots the variations in the diagrams
"""
def plot_variations(X,X_orig,y_train, sign_name,num_var,num_examples = 5 ):
fig, axes = plt.subplots(nrows=num_examples, ncols=num_var+1)
index_list = []
sign_name_list =[]
pad = 5 # in points
for j in range(1,num_examples+1) :
y_index = random.randint(0,len(y_train))
index_list.append(y_index)
sign_name_list.append(sign_name[y_train[y_index]])
ax_lbls = axes[:,0]
for ax, row in zip(ax_lbls, sign_name_list):
ax.annotate(row.replace(" ","\n"), xy=(0, 0.5), xytext=(-ax.yaxis.labelpad - pad, 0),
xycoords=ax.yaxis.label, textcoords='offset points',
size='small', ha='right', va='center')
#ax.set_ylabel(row.replace(" ","\n"), rotation=0, size='small')
for j in range(1,num_examples+1) :
y_index = index_list[j-1]
image_index = y_index*num_var
ax1 = axes[j-1,0]
#ax1 = plt.subplot(num_examples,num_var,k)
ax1.imshow(X_orig[y_index])
ax1.get_xaxis().set_ticks([])
ax1.get_yaxis().set_ticks([])
for i in range(1,num_var +1):
k = (num_var*(j-1))+i
im_index = int(str(num_examples)+str(num_var)+str(k))
ax1 = axes[j-1,i]
#ax1 = plt.subplot(num_examples,num_var,k)
ax1.imshow(X[image_index+i-1])
ax1.get_xaxis().set_ticks([])
ax1.get_yaxis().set_ticks([])
#ax1.set_ylabel(sign_name[y_train[y_index]], rotation=0, size='small')
fig.tight_layout()
"""
Converts the array of images into gray scale images
"""
def gray_scale(X):
X_gray = []
for each in X:
#print(np.shape(each))
gray_image = np.asarray(cv2.cvtColor(each, cv2.COLOR_RGB2GRAY))
#print(type(gray_image))
X_gray.append(np.expand_dims(gray_image,axis=2))
#plt.imshow(X_gray[0])
return X_gray
# Rotation of image
"""
Rotate the image clockwise and counter clock wise. [-6,-4,-2,0,2,4,6]
6 images per image.
"""
def image_rotation(X,y):
rot_X = []
rot_y = []
for ex,ey in zip(X,y):
img = Image.fromarray(ex)
i = random.randrange(-10,10)
im2 = img.convert('RGBA')
new_image = im2.rotate(i)
# a white image same size as rotated image
blank_img = Image.new('RGBA', (32,32), (255,)*4)
# create a composite image using the alpha layer of rot as a mask
out = Image.composite(new_image, blank_img, new_image)
out.convert('RGB')
rot_X.append(np.asarray(out))
rot_y.append(ey)
return rot_X,rot_y
"""
Blur the image with a kernel of 5x5
"""
def image_blur(X,y,kernel_size=5):
blur_X =[]
blur_y = []
for ex,ey in zip(X,y):
#kernel = np.ones((kernel_size,kernel_size),np.float32)/25
blr_img = cv2.GaussianBlur(ex,(kernel_size,kernel_size),0)
#blr_img = cv2.filter2D(ex,-1,kernel)
blur_X.append(blr_img)
blur_y.append(ey)
return blur_X,blur_y
def noisy_data(X,y):
row,col,ch = np.shape(X[0])
noisy_X = []
noisy_y = []
for ex,ey in zip(X,y):
noisy_image = np.asarray(
np.random.uniform(
-5,5,(row,col,ch))
,dtype=np.uint8)+ex
noisy_X.append(noisy_image)
noisy_y.append(ey)
return noisy_X,noisy_y
"""
Contrast Increase using HSV values.
"""
def contrast_increase_data(X,y,value=5):
row,col,ch = np.shape(X[0])
contra_X = []
contra_y = []
for ex,ey in zip(X,y):
#contra_image = np.array(cv2.cvtColor(ex,cv2.COLOR_RGB2YUV),dtype=np.float32)
im2 = cv2.cvtColor(ex, cv2.COLOR_RGB2GRAY)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
X_ = clahe.apply(im2)
contra_X.append(X_)
contra_y.append(ey)
return contra_X,contra_y
def transform_data(X,y,count=1000):
ang_range = 20
trans_range = 10
shear_range = 5
X_new = []
y_new = []
for ex,ey in zip(X,y):
ang_rot = np.random.uniform(ang_range)-ang_range/2
rows,cols,ch = np.shape(ex)
Rot_M = cv2.getRotationMatrix2D((cols/2,rows/2),ang_rot,1)
tr_x = trans_range*np.random.uniform()-trans_range/2
tr_y = trans_range*np.random.uniform()-trans_range/2
Trans_M = np.float32([[1,0,tr_x],[0,1,tr_y]])
pts1 = np.float32([[5,5],[20,5],[5,20]])
pt1 = 5+shear_range*np.random.uniform()-shear_range/2
pt2 = 20+shear_range*np.random.uniform()-shear_range/2
pts2 = np.float32([[pt1,5],[pt2,pt1],[5,pt2]])
shear_M = cv2.getAffineTransform(pts1,pts2)
img = cv2.warpAffine(ex,Rot_M,(cols,rows))
img = cv2.warpAffine(img,Trans_M,(cols,rows))
img = cv2.warpAffine(img,shear_M,(cols,rows))
X_new.append(img)
y_new.append(ey)
return X_new,y_new
Preprocessing set output (Contrast increase and gray scale)¶
sign_name = list(data.SignName)
X_rot,y_rot = contrast_increase_data(X_train,y_train,value=200)
X_new = X_rot
X_train_gray = np.reshape(gray_scale(X_train),(len(X_train),32,32))
plot_variations(np.reshape(X_new,(len(X_new),32,32)),X_train_gray,y_train,sign_name,1,5)
Step 2: Design and Test a Model Architecture¶
Design and implement a deep learning model that learns to recognize traffic signs. Train and test your model on the German Traffic Sign Dataset.
There are various aspects to consider when thinking about this problem:
- Neural network architecture
- Play around preprocessing techniques (normalization, rgb to grayscale, etc)
- Number of examples per label (some have more than others).
- Generate fake data.
Here is an example of a published baseline model on this problem. It's not required to be familiar with the approach used in the paper but, it's good practice to try to read papers like these.
NOTE: The LeNet-5 implementation shown in the classroom at the end of the CNN lesson is a solid starting point. You'll have to change the number of classes and possibly the preprocessing, but aside from that it's plug and play!
Implementation¶
Use the code cell (or multiple code cells, if necessary) to implement the first step of your project. Once you have completed your implementation and are satisfied with the results, be sure to thoroughly answer the questions that follow.
Modified LetNet for traffic Sign classifier¶
def transform_image(X,y):
ang_range = 20
trans_range = 10
shear_range = 5
ang_rot = np.random.uniform(ang_range)-ang_range/2
rows,cols,ch = np.shape(X)
Rot_M = cv2.getRotationMatrix2D((cols/2,rows/2),ang_rot,1)
tr_x = trans_range*np.random.uniform()-trans_range/2
tr_y = trans_range*np.random.uniform()-trans_range/2
Trans_M = np.float32([[1,0,tr_x],[0,1,tr_y]])
pts1 = np.float32([[5,5],[20,5],[5,20]])
pt1 = 5+shear_range*np.random.uniform()-shear_range/2
pt2 = 20+shear_range*np.random.uniform()-shear_range/2
pts2 = np.float32([[pt1,5],[pt2,pt1],[5,pt2]])
shear_M = cv2.getAffineTransform(pts1,pts2)
img = cv2.warpAffine(X,Rot_M,(cols,rows))
img = cv2.warpAffine(img,Trans_M,(cols,rows))
img = cv2.warpAffine(img,shear_M,(cols,rows))
return img,y
def augment_data(X_train,y_train,class_min_count=1000):
MIN_CLASS_LEN=class_min_count
X_final = []
y_final = []
for i in range(43):
indexes = np.where(y_train==i)
idx = np.reshape(indexes, (np.shape(indexes)[1]))
len_class = np.shape(indexes)[1]
X_class = X_train[idx]
#print(np.shape(X_class))
#plt.imshow(X_class[0],cmap="gray")
y_class = np.asarray(y_train)[idx]
X_class_new = np.asarray(np.copy(X_class))
y_class_new = np.copy(y_class)
if(len_class<MIN_CLASS_LEN):
#print(len_class)
for i in range(0,MIN_CLASS_LEN-len_class):
index = random.randint(0,len_class-1)
new_image,y = transform_image(X_class[index],y_class[index])
X_final.append(new_image)
#print(y)
y_final.append(y)
for ex,ey in zip(X_class_new,y_class_new):
X_final.append(ex)
y_final.append(ey)
return X_final,y_final
X_final,y_final = augment_data(X_train,y_train)
#print(len(y_final))
After augmentation of data (Adding more data to the classes with less data)¶
sns.distplot(y_final,kde=False)
X_train_contrast,y_train_contrast = contrast_increase_data(X_final,y_final)
X_test_final,y_test_final = contrast_increase_data(X_test,y_test)
X_test_final = np.asarray(np.expand_dims(np.asarray(X_test_final) ,axis=3))
X_train_augmented = np.asarray(np.expand_dims(np.asarray(X_train_contrast) ,axis=3))
y_train_augmented = np.expand_dims(y_train_contrast,axis=1)
from sklearn.utils import shuffle
X,y = shuffle(X_train_augmented,y_train_augmented)
import math
X_validation_final = X[:math.ceil(len(y)/10):,:,]
y_validation_final = y[:math.ceil(len(y)/10)]
X_train_final = X[math.ceil(len(y)/10):,:,]
y_train_final = y[math.ceil(len(y)/10):]
y_train_final = np.reshape(y_train_final,np.shape(y_train_final)[0])
y_validation_final = np.reshape(y_validation_final,np.shape(y_validation_final)[0])
sns.distplot(y_train_final,kde=False)
print(np.shape(X_validation_final))
print(np.shape(X_train_final))
print(np.shape(X_train_augmented))
print(np.shape(y_validation_final))
print(np.shape(y_train_final))
print(np.shape(y_train_augmented))
plt.imshow(np.reshape(X_test_final[23],(32,32)),cmap="gray")
print(np.shape(y_test_final))
y_train_final[23]
X_new = X_train_final#.astype(float)
X_new = X_new.astype(float)
m = np.mean(X_train_final)
st = np.std(X_train_final)
X_new-=m
X_train_final = X_new /255.0
X_val = X_validation_final#.astype(float)
X_val = X_val.astype(float)
X_val-=m
X_validation_final = X_val /255.0
X_test = X_test_final#.astype(float)
X_test = X_test.astype(float)
X_test-=m
X_test_final = X_test /255.0
### Preprocess the data here.
### Feel free to use as many code cells as needed.
from tensorflow.contrib.layers import flatten
import tensorflow as tf
def conv2d(x,weight,bias,strides=1,padding='SAME'):
layer = tf.nn.conv2d(x,weight,[1,strides,strides,1],padding=padding)
layer = tf.nn.bias_add(layer,bias)
layer = tf.nn.relu(layer)
return layer
def maxpool(x,k=2,padding='SAME'):
kernel_size = [1,k,k,1]
return tf.nn.max_pool(x,ksize=kernel_size,strides=[1,k,k,1],padding=padding)
def fully_connected(in_tensor,layers,keep_prob):
mu = 0
sigma = 0.1
in_dim = np.product(in_tensor.get_shape().as_list()[1:])
tensor_flatten = flatten(in_tensor)
print(in_dim)
count = 1
for each in layers:
with tf.name_scope("fully-connected-"+str(count)):
weights = tf.Variable(tf.random_normal([in_dim, each],mu,sigma,dtype=tf.float32))
bias = tf.Variable(tf.zeros(each))
fc1_drop = tf.nn.dropout(tensor_flatten, keep_prob)
fc2 = tf.add(
tf.matmul(fc1_drop, weights),
bias)
# TODO: Activation.
tensor_flatten = tf.nn.tanh(fc2)
in_dim = each
count = count +1
print(in_dim)
return tensor_flatten
import tensorflow.contrib.slim as slim
def LeNet_multi_batchnorm(x,keep_prob):
# Hyperparameters
mu = 0
sigma = 0.1
with tf.name_scope('CNN'):
with tf.name_scope('conv1'):
input_layer = slim.conv2d(x,40,[3,3],activation_fn=None)
with tf.name_scope('batch-norm'):
input_layer_1 = slim.batch_norm(input_layer,activation_fn=None)
with tf.name_scope('relu'):
input_layer_1 = tf.nn.relu(input_layer_1)
with tf.name_scope('maxpool'):
layer_1 = maxpool(input_layer_1,k=2)
print(layer_1.get_shape())
with tf.name_scope('conv2'):
input_layer_2 = slim.conv2d(layer_1,60,[3,3],activation_fn=None)
with tf.name_scope('batch-norm'):
input_layer_2 = slim.batch_norm(input_layer_2,activation_fn=None)
with tf.name_scope('relu'):
input_layer_2 = tf.nn.relu(input_layer_2)
with tf.name_scope('maxpool'):
layer_2 = maxpool(input_layer_2,k=2)
with tf.name_scope('dropout'):
layer_2 = tf.nn.dropout(layer_2, keep_prob)
print(layer_2.get_shape())
with tf.name_scope('conv3'):
input_layer_3 = slim.conv2d(layer_2,125,[3,3],activation_fn=None)
with tf.name_scope('batch-norm'):
input_layer_3 = slim.batch_norm(input_layer_3,activation_fn=None)
with tf.name_scope('relu'):
input_layer_3 = tf.nn.relu(input_layer_3)
with tf.name_scope('maxpool'):
layer_3 = maxpool(input_layer_3,k=2)
with tf.name_scope('dropout'):
layer_3 = tf.nn.dropout(layer_3, keep_prob)
print(layer_3.get_shape())
fc1 = flatten(layer_3)
# fc1 = tf.concat(1,(fc1,flatten(layer_1)))
with tf.name_scope('fully-connected'):
logits = fully_connected(fc1,[512,512,43],keep_prob)
return logits
INPUT_DEPTH = 1
tf.reset_default_graph()
x = tf.placeholder(tf.float32, (None, 32, 32, INPUT_DEPTH))
y = tf.placeholder(tf.int32, (None))
keep_prob = tf.placeholder(tf.float32)
one_hot_y = tf.one_hot(y, 43)
BATCH_SIZE = 256
import tensorflow as tf
from sklearn.utils import shuffle
from sklearn.model_selection import train_test_split
rate = 0.001
logits = LeNet_multi_batchnorm(x,keep_prob)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, one_hot_y)
loss_operation = tf.reduce_mean(cross_entropy)
optimizer = tf.train.AdamOptimizer(learning_rate = rate)
# Gradient clipping
gvs = optimizer.compute_gradients(loss_operation)
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs]
optimizer.apply_gradients(capped_gvs)
training_operation = optimizer.minimize(loss_operation)
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(one_hot_y, 1))
accuracy_operation = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
saver = tf.train.Saver()
def evaluate(X_data, y_data):
num_examples = len(X_data)
total_accuracy = 0
sess = tf.get_default_session()
for offset in range(0, num_examples, BATCH_SIZE):
batch_x, batch_y = X_data[offset:offset+BATCH_SIZE], y_data[offset:offset+BATCH_SIZE]
accuracy = sess.run(accuracy_operation, feed_dict={x: batch_x, y: batch_y,keep_prob:1.0})
total_accuracy += (accuracy * len(batch_x))
return total_accuracy / num_examples
np.shape(y_train_final)
import os
import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"
EPOCHS = 100
model_file = "./final_version_review_v4"
# model_file = "./lenet_multi_batch_norm_final_v5"
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
num_examples =len(X_train)#392090
from pathlib import Path
my_file = Path(model_file+".meta")
if my_file.is_file():
saver.restore(sess,model_file)
print("Model restored...")
print("Training...")
print()
X_train_1 = X_train_final
y_train_1 = y_train_final
train_writer = tf.train.SummaryWriter('./train',sess.graph)
print("Total "+str(len(y_train_1)))
for i in range(EPOCHS):
X_train, y_train = shuffle(X_train_1, y_train_1)
acc = []
ent = []
for offset in range(0, num_examples, BATCH_SIZE):
end = offset + BATCH_SIZE
batch_x, batch_y = np.copy(X_train[offset:end]), np.copy(y_train[offset:end])
sess.run(training_operation, feed_dict={x: batch_x, y: batch_y,keep_prob:0.8})
training_accuracy = evaluate(batch_x, batch_y)
mean_cross_entropy = sess.run(loss_operation, feed_dict={x: batch_x, y: batch_y,keep_prob:1.0})
acc.append(training_accuracy)
ent.append(mean_cross_entropy)
validation_accuracy = evaluate(X_validation_final, y_validation_final)
print("EPOCH {} ...".format(i+1))
print("Training Accuracy = {:.3f}".format(np.mean(acc)))
print("Cross entropy = {:.4f}".format(np.mean(ent)))
print("Validation Accuracy = {:.3f}".format(validation_accuracy))
if(i%10==0):
saver.save(sess, model_file)
print("Model saved")
test_accuracy = evaluate(X_test, y_test_final)
print("Test Accuracy = {:.3f}".format(test_accuracy))
saver.save(sess, model_file)
print("Model saved")
Test Accuracy : 96.9%¶
# Reducing the Entropy running more iterations
rate = 0.001 #fine tune with smaller rate
EPOCHS = 100
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
num_examples =len(X_train_final)#392090
from pathlib import Path
model_save = "final_finetuning_v2"
my_file = Path(model_file+".meta")
if my_file.is_file():
saver.restore(sess,model_file)
print("Model restored...")
print("Training...")
X_train_1 = X_train_final
y_train_1 = y_train_final
train_writer = tf.train.SummaryWriter('./train',sess.graph)
print("Total "+str(len(y_train_1)))
for i in range(EPOCHS):
X_train, y_train = shuffle(X_train_1, y_train_1)
acc = []
ent = []
for offset in range(0, num_examples, BATCH_SIZE):
end = offset + BATCH_SIZE
batch_x, batch_y = np.copy(X_train[offset:end]), np.copy(y_train[offset:end])
sess.run(training_operation, feed_dict={x: batch_x, y: batch_y,keep_prob:0.75})
training_accuracy = evaluate(batch_x, batch_y)
mean_cross_entropy = sess.run(loss_operation, feed_dict={x: batch_x, y: batch_y,keep_prob:1.0})
acc.append(training_accuracy)
ent.append(mean_cross_entropy)
validation_accuracy = evaluate(X_validation_final, y_validation_final)
print("EPOCH {} ...".format(i+1))
print("Training Accuracy = {:.3f}".format(np.mean(acc)))
print("Cross entropy = {:.4f}".format(np.mean(ent)))
print("Validation Accuracy = {:.3f}".format(validation_accuracy))
if(i%10==0):
saver.save(sess, model_save)
print("Model saved")
test_accuracy = evaluate(X_test, y_test_final)
print("Test Accuracy = {:.3f}".format(test_accuracy))
saver.save(sess, model_save)
print("Model saved")
Test Accuracy of 97.5¶
Question 1¶
Describe how you preprocessed the data. Why did you choose that technique?
Answer:
Preprocessing steps : Gray scale conversion and contrast increase of the images are the two preprocessing that was done.
Even though the color gives us some indication on the type of sign, but training a grayscale image was much faster, with probably slight decrease in the performance. Initially, started off with the RGB images, but grayscale perforcewell as well. Also, when increased the contrast, some of the features in the image became much clearer.
Also, Normalized the data by subtracting the mean and dividing the values by 255 ( Max pixel value ),so that this gives stability in the weights that is learnt. Along with this, I have also used Gradient clipping so that the value does not go beyond -1 and 1. And have also used batch normalization at multiple layers (Layer wise, batchwise normalization, not just at the input level)
Here are the few things that were done :
- Grayscale image
- Contrast-limited Adaptive Histogram Equalization (CLAHE)
- Normalized the image to fall in between (-1 and 1)
### Generate data additional data (OPTIONAL!)
# With slight rotation only as flipping reverses the direction, so avoided some of the data preprocessing steps.
# Contrast increase
#
### and split the data into training/validation/testing sets here.
### Feel free to use as many code cells as needed.
Question 2¶
Describe how you set up the training, validation and testing data for your model. Optional: If you generated additional data, how did you generate the data? Why did you generate the data? What are the differences in the new dataset (with generated data) from the original dataset?
Answer:
Data was generated using the following methods
- Adding noise ( To make CNN roboust to noise , Note : was not used in the final submission, This added high bias, might have to be careful how much to use)
- Bluring of the image ( Some were used initial training. This plus few other trick took the model from 94 % validation accuracy to 96% )
- Lot more data was generated as some of the classes had liited set of data ( Rotation, Shear, Zoom etc was used). These techniques were from the fourm discussions and helper methods.
### Define your architecture here.
### Feel free to use as many code cells as needed.
Question 3¶
What does your final architecture look like? (Type of model, layers, sizes, connectivity, etc.) For reference on how to build a deep neural network using TensorFlow, see Deep Neural Network in TensorFlow from the classroom.
Answer:
Model consists of the following layers
3 Conv layers with depth 40 , 60 and 125 with 3x3 filter , stride 1
Alternating each of these layers is a 2x2 Maxpool with a stride of 2
Batch normalization is done after relu in each of these conv,maxpool pair
2 Fully connected layer of size 512 with dropouts layer.
Gradient clipping is done across all the parameters (-1,1)
### Train your model here.
### Feel free to use as many code cells as needed.
# Architecture diagram + Tensorflow graph.
img = Image.open("./CNN_graph.png")
img
img = Image.open("./CNN_components.png")
img
Question 4¶
How did you train your model? (Type of optimizer, batch size, epochs, hyperparameters, etc.)
Answer:
- Optimizer : Adam
- EPOCS tried : For some of the larger models that I tried upto 200 EPOCS , but smaller models would fit within 50 EPOCS.
- Batch size : 128,256,512 and 1024 was tried , 256 seems to be the best for the current model.
- learning rate : 0.01 to 0.0001 was tried, higher learning rate would diverge and lower would take its own time to coverge, 0.001 was the best for the model ( Coverges within 100 EPOCS ).
Question 5¶
What approach did you take in coming up with a solution to this problem? It may have been a process of trial and error, in which case, outline the steps you took to get to the final solution and why you chose those steps. Perhaps your solution involved an already well known implementation or architecture. In this case, discuss why you think this is suitable for the current problem.
Answer: Things tried
- Augment Data with -5 degree to 5 degree purterbation
- Gray scale (Image convert)
- Image Normalization
- LeNet Architecture
- Adding the drop out layer for the fully connected piece
- Simpler model for Gray scale ( 94 % validation ). Removing the one layer of fully connected.
- Layers were added to the LeNet architecture
- ResNet, Multi-scale and Inception Models were tried, but the convergence was very slow compared to simpler models, But could give better result if trained well.
Data Augmentation was done: Following things were tried
- Noise to te input data
- Blurring of the input data
- Rotation of the data (-15 to 15 )
- Ideas from the fourm were incorporated as well, shear, rotation and
When it was Overfitting following things were added
- Regularizer/ Gradient clipping.
- Batch Normalization
Following things are yet to be tried:
- Fine tune of the model using even more data augmentation
Initially took trial & error approach but later Roughly, following method was used to train the models ( example error rate )
| Human Error | Training Error | Validation Error |
|---|---|---|
| 1% | 5% | 6% |
| Human Error | Training Error | Validation Error |
|---|---|---|
| 1% | 2% | 10% |
Reference:¶
Andrew Ng's Presentation on nuts and bolts of deep learning : https://www.youtube.com/watch?v=F1ka6a13S9I
Step 3: Test a Model on New Images¶
Take several pictures of traffic signs that you find on the web or around you (at least five), and run them through your classifier on your computer to produce example results. The classifier might not recognize some local signs but it could prove interesting nonetheless.
You may find signnames.csv useful as it contains mappings from the class id (integer) to the actual sign name.
Implementation¶
Use the code cell (or multiple code cells, if necessary) to implement the first step of your project. Once you have completed your implementation and are satisfied with the results, be sure to thoroughly answer the questions that follow.
### Load the images and plot them here.
### Feel free to use as many code cells as needed.
# Load pickled data
import pickle
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
import pandas as pd
import random
from PIL import Image
import pickle
import numpy as np
import cv2
import tensorflow as tf
import matplotlib.pyplot as plt
# Visualizations will be shown in the notebook.
%matplotlib inline
from os import listdir
from os.path import isfile, join
def extra_data():
training_dirs =[]
l = []
dirs = listdir("Training/")
for each in dirs:
if( not each.endswith("txt")):
training_dirs.append(each)
for each in training_dirs:
dirs_inner = listdir("Training/"+each)
for e in dirs_inner:
if(e.endswith(".ppm")):
l.append(join("Training",each,e))
return l
def image_resize(images):
image_list = []
for i in images:
img = Image.open(i)
img = img.resize((32, 32), Image.ANTIALIAS)
image_list.append(img)
return image_list
l = extra_data()
"""
Plots the images
"""
def plot_images(X,n_rows,n_cols ):
fig, axes = plt.subplots(nrows=n_rows, ncols=n_cols)
for i in range(n_rows):
for j in range(n_cols) :
ax1 = axes[i,j]
#ax1 = plt.subplot(num_examples,num_var,k)
ax1.imshow(X[n_cols*i+j])
ax1.get_xaxis().set_ticks([])
ax1.get_yaxis().set_ticks([])
plt.show()
fig.tight_layout()
allowed_list = ["00031","00021","00006","00007","00057","00004",""]
new_list = []
for e in l :
clas = e.split("/")[1]
if(clas in allowed_list):
new_list.append(e)
import random
SAMPLE_SIZE = 10
ims = random.sample(new_list,SAMPLE_SIZE)
X_new_imgs = image_resize(ims)
plot_images(X_new_imgs,2,5)
Reference:¶
J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. The German Traffic Sign Recognition Benchmark: A multi-class classification competition. In Proceedings of the IEEE International Joint Conference on Neural Networks, pages 1453–1460. 2011.
Question 6¶
Choose five candidate images of traffic signs and provide them in the report. Are there any particular qualities of the image(s) that might make classification difficult? It could be helpful to plot the images in the notebook.
Answer:
### Run the predictions here.
### Feel free to use as many code cells as needed.
# X_train_contrast_new,y_train_contrast_new = contrast_increase_data(np.asarray(X_new_imgs),np.ones(SAMPLE_SIZE))#dummy)
# X_train_augmented_new = np.asarray(np.expand_dims(np.asarray(X_train_contrast_new) ,axis=3))
im_array = [np.asarray(im) for im in X_new_imgs]
X_train_contrast_new,y_train_contrast_new = contrast_increase_data(np.asarray(im_array),np.ones(SAMPLE_SIZE))#dummy)
X_train_new_images = np.asarray(np.expand_dims(np.asarray(X_train_contrast_new) ,axis=3))
model_file = "./final_finetuning_v2"
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
num_examples =len(X_train_new_images)#392090
from pathlib import Path
my_file = Path(model_file+".meta")
if my_file.is_file():
saver.restore(sess,model_file)
logit_out = sess.run(logits,feed_dict ={x:X_train_new_images,keep_prob:1.0})
softmax_out = sess.run(tf.nn.softmax(logit_out))
n_rows = 10
n_cols = 1
for i in range(n_rows):
plt.plot(figsize=(20, 25))
#ax1 = axes#[i]
plt.imshow(X_new_imgs[i], aspect='auto', extent=(0.2, 12.0, .5, .7), zorder=-1)
#ax1.get_yaxis().set_ticks([])
ind = range(len(softmax_out[i]))
plt.xticks()
plt.xticks(np.arange(0.5,43.5,1.0), list(data.SignName),ha='right',rotation=45)
rects1 = plt.bar(ind, softmax_out[i])
st = data.SignName[int(np.argmax(softmax_out[i]))]
plt.title("Predicted: "+st)
plt.show()
Observation¶
- Eventhough the accuracy of the model is high (97.5% ). The entropy seems to be high as well, hence the confidence on prediction is not that high.
- Some of the signs that are not in the training data are pridicted wrong. Bicycle , get confused to stop (Note we do not use color)
- Some of the ones it gets wrong are Ahead only. Also, there is confusion between "No passing" and "No entry", As the two vehicles that are close together might get confused for no entry.
Question 7¶
Is your model able to perform equally well on captured pictures when compared to testing on the dataset? The simplest way to do this check the accuracy of the predictions. For example, if the model predicted 1 out of 5 signs correctly, it's 20% accurate.
NOTE: You could check the accuracy manually by using signnames.csv (same directory). This file has a mapping from the class id (0-42) to the corresponding sign name. So, you could take the class id the model outputs, lookup the name in signnames.csv and see if it matches the sign from the image.
Answer: Accuracy on the captured images is 70% while it was 97.5% on the testing set. Models seems to be overfitting on the test set.
### Visualize the softmax probabilities here.
### Feel free to use as many code cells as needed.
Question 8¶
Use the model's softmax probabilities to visualize the certainty of its predictions, tf.nn.top_k could prove helpful here. Which predictions is the model certain of? Uncertain? If the model was incorrect in its initial prediction, does the correct prediction appear in the top k? (k should be 5 at most)
tf.nn.top_k will return the values and indices (class ids) of the top k predictions. So if k=3, for each sign, it'll return the 3 largest probabilities (out of a possible 43) and the correspoding class ids.
Take this numpy array as an example:
# (5, 6) array
a = np.array([[ 0.24879643, 0.07032244, 0.12641572, 0.34763842, 0.07893497,
0.12789202],
[ 0.28086119, 0.27569815, 0.08594638, 0.0178669 , 0.18063401,
0.15899337],
[ 0.26076848, 0.23664738, 0.08020603, 0.07001922, 0.1134371 ,
0.23892179],
[ 0.11943333, 0.29198961, 0.02605103, 0.26234032, 0.1351348 ,
0.16505091],
[ 0.09561176, 0.34396535, 0.0643941 , 0.16240774, 0.24206137,
0.09155967]])Running it through sess.run(tf.nn.top_k(tf.constant(a), k=3)) produces:
TopKV2(values=array([[ 0.34763842, 0.24879643, 0.12789202],
[ 0.28086119, 0.27569815, 0.18063401],
[ 0.26076848, 0.23892179, 0.23664738],
[ 0.29198961, 0.26234032, 0.16505091],
[ 0.34396535, 0.24206137, 0.16240774]]), indices=array([[3, 0, 5],
[0, 1, 4],
[0, 5, 1],
[1, 3, 5],
[1, 4, 3]], dtype=int32))Looking just at the first row we get [ 0.34763842, 0.24879643, 0.12789202], you can confirm these are the 3 largest probabilities in a. You'll also notice [3, 0, 5] are the corresponding indices.
Answer: Cross entropy seems to be high as it seems to give small probabilities to each class.
Children crossing seems to get confused with pedestrians.( But Children crossing seems to be in the top 5)
Bicycle crossing gets confused with pedestrians, Children crossing , Beware of ice/Snow. In one of the images it does not pridict in top5.
Top5 prediction accuracy is 80% compared to 70% in top 1 for these 10 new test images.
Note: Once you have completed all of the code implementations and successfully answered each question above, you may finalize your work by exporting the iPython Notebook as an HTML document. You can do this by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
num_examples =len(X_train_new_images)#392090
from pathlib import Path
my_file = Path(model_file+".meta")
if my_file.is_file():
saver.restore(sess,model_file)
logit_out = sess.run(logits,feed_dict ={x:X_train_new_images,keep_prob:1.0})
softmax_out = sess.run(tf.nn.softmax(logit_out))
prediction = sess.run(tf.nn.top_k(softmax_out, k=5, sorted=True))
p,c = prediction
for i in range(10):
plt.plot(figsize=(5, 5))
#ax1 = axes#[i]
plt.imshow(X_new_imgs[i], aspect='auto', extent=(0.2, 4.0, .5, .7), zorder=-1)
#ax1.get_yaxis().set_ticks([])
ind = range(len(p[i]))
#print(c[i])
lbls = list(data.SignName)
top_5lbl = [lbls[j] for j in c[i] ]
plt.xticks(np.arange(0.5,5.5,1.0),top_5lbl ,ha='right',rotation=45)
rects1 = plt.bar(ind, p[i])
st = data.SignName[c[i][0]]
plt.title("Predicted: "+st)
plt.show()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
num_examples =len(X_test)#392090
from pathlib import Path
my_file = Path(model_file+".meta")
if my_file.is_file():
saver.restore(sess,model_file)
for i in range(43):
indexes = np.where(np.asarray(y_test_final)==i)
#print(len(indexes[0]))
idx = np.reshape(indexes, (np.shape(indexes)[1]))
len_class = np.shape(indexes)[1]
X_class = X_test_final[idx]
logit_out = sess.run(tf.nn.softmax(logits), feed_dict ={x:X_class,keep_prob:1.0})
prediction = sess.run(tf.nn.top_k(logit_out, k=5, sorted=True))
#print(prediction[1])
accurate = 0
#print(prediction)
for prob in prediction[1]:
if i in prob:
accurate=accurate+1
print("Top 5 Accuracy on Test data for ", data.SignName[i], " : ",np.float(accurate)/len(indexes[0]))